NanoForecast v0.5: Competitive Time Series Forecasting Through Training Pipeline Optimization
NanoForecast v0.5, a 6.5M forecaster, beats much larger TimesFM on ETT after pipeline fixes.
NanoForecast v0.5 is a 6.5-million-parameter time-series forecaster that, after training-pipeline fixes and no architecture change, competes with TimesFM at 200 million parameters. Corrected loss-scope handling, tensor-shape alignment, and wider augmentation cut overall MASE by 43.8% (3.030 to 1.704) under one protocol and the same data and compute. It beats TimesFM on all three ETT sets (MASE 0.676/1.110/0.287 versus 0.705/1.360/0.545) and on exchange rate (4.317 versus 4.383), while TimesFM leads on electricity and traffic; it also beats official PatchTST on all three ETT sets. Training takes about 12 hours on an NVIDIA T4, CPU inference was measured on an Apple M4, and code, checkpoints, and the eval framework are released under Apache 2.0.
- A 6.5M model matches or beats 200M TimesFM on ETT and exchange-rate sets.
- Pipeline fixes cut overall MASE 43.8%, from 3.030 to 1.704, with no architecture change.
- TimesFM still leads on high-cardinality electricity and traffic datasets.
- Training is about 12 hours on one NVIDIA T4; weights are Apache 2.0.
- AI models
- NanoForecast v0.5TimesFMPatchTST
From the original article
1 min read at huggingface.co
We present NanoForecast v0. 5, a 6. 5M-parameter forecaster that competes with models 31x its size (TimesFM, 200M parameters) after training pipeline fixes and no architecture change.
By Gautam Kishore · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.